Fig. 2: Benchmarking techniques and metrics to assess optimisation performance.

a Four Suzuki coupling virtual datasets from Olympus³³, derived from experimental data, are used for benchmarking in this study. b The C-H arylation virtual dataset is generated by training a machine learning (ML) model on 1728 experimentally collected reactions from Torres et al.¹⁴ (EDBO+), then predicting reaction outcomes for a larger range of reaction conditions and variables not present in the original training data. This creates a large-scale virtual dataset suitable for benchmarking high-throughput experimentation (HTE) optimisation campaigns (see Supplementary Information Section 1 and Methods). c Distribution of reaction objectives, yield (%) and catalyst turnover number, for the first Suzuki Coupling (i) virtual dataset from Olympus³³. Pareto points represent the optimal multi-objective combinations. d Distribution of reaction objectives, yield (%) and reaction cost, for the C-H arylation virtual dataset generated in this study. e The hypervolume is used to assess the performance of optimisation algorithms. The hypervolume quantifies the volume enclosed by the optimal objective (e.g., yield-selectivity) combinations (Pareto points) identified by each algorithm. The hypervolume is used to compare algorithm performance against the best reaction conditions in the benchmark dataset, evaluating how effectively the best existing solutions are identified.